Evaluating Chess Moves by Analysing Sentiments in Teaching Textbooks

Abstract

The rules of playing chess are simple to comprehend, and yet it is challenging to make accurate decisions in the game. Hence, chess lends itself well to the development of an artificial intelligence (AI) system that simulates real-life problems, such as in decision-making processes. Learning chess strategies has been widely investigated, with most studies focused on learning from previous games using search algorithms. Chess textbooks encapsulate grandmaster knowledge, which explains playing strategies. This thesis investigates three research questions on the possibility of unlocking hidden knowledge in chess teaching textbooks. Firstly, we contribute to the chess domain with a new heterogeneous chess dataset “LEAP”, consists of structured data that represents the environment “board state”, and unstructured data that represent explanation of strategic moves. Additionally, we build a larger unstructured synthetic chess dataset to improve large language models familiarity with the chess teaching context. With the LEAP dataset, we examined the characteristics of chess teaching textbooks and the challenges of using such a data source for training Natural Language (NL)-based chess agent. We show by empirical experiments that following the common approach of sentence-level evaluation of moves are not insightful. Secondly, we observed that chess teaching textbooks are focused on explanation of the move’s outcome for both players alongside discussing multiple moves in one sentence, which confused the models in move evaluation. To address this, we introduce an auxiliary task by using verb phrase-level to evaluate the individual moves. Furthermore, we show by empirical experiments the usefulness of adopting the Aspect-based Sentiment Analysis (ABSA) approach as an evaluation method of chess moves expressed in free-text. With this, we have developed a fine-grained annotation and a small-scale dataset for the chess-ABSA domain “ASSESS”. Finally we examined the performance of a fine-tuned LLM encoder model for chess-ABSA and showed that the performance of the model for evaluating chess moves is comparable to scores obtained from a chess engine, Stockfish. Thirdly, we developed an instruction-based explanation framework, using prompt engineering with zero-shot learning to generate an explanation text of the move outcome. The framework also used a chess ABSA decoder model that uses an instructions format and evaluated its performance on the ASSESS dataset, which shows an overall improvement performance. Finally, we evaluate the performance of the framework and discuss the possibilities and current challenges of generating large-scale unstructured data for the chess, and the effect on the chess-ABSA decoder model.